Physiological Time Series Data Research Articles

Concept-based AI interpretability in physiological time-series data: Example of abnormality detection in electroencephalography

Background and ObjectiveDespite recent performance advancements, deep learning models are not yet adopted in clinical practice on a wide scale. The intrinsic intransparency of such systems is commonly cited as one major reason for this reluctance. This has motivated methods that aim to provide explanations of model functioning. Known limitations of feature-based explanations have led to an increased interest in concept-based interpretability. Testing with Concept Activation Vectors (TCAV) employs human-understandable, abstract concepts to explain model behavior. The method has previously been applied to the medical domain in the context of electronic health records, retinal fundus images and magnetic resonance imaging. MethodsWe explore the usage of TCAV for building interpretable models on physiological time series, using an example of abnormality detection in electroencephalography (EEG). For this purpose, we adopt the XceptionTime model, which is suitable for multi-channel physiological data of variable sizes. The model provides state-of-the-art performance on raw EEG data and is publically available. We propose and test several ideas regarding concept definition through metadata mining, using additional labeled EEG data and extracting interpretable signal characteristics in the form of frequencies. By including our own hospital data with analog labeling, we further evaluate the robustness of our approach. ResultsThe tested concepts show a TCAV score distribution that is in line with the clinical expectations, i.e. concepts known to have strong links with EEG pathologies (such as epileptiform discharges) received higher scores than the neutral concepts (e.g. sex). The scores were consistent across the applied concept generation strategies. ConclusionsTCAV has the potential to improve interpretability of deep learning applied to multi-channel signals as well as to detect possible biases in the data. Still, further work on developing the strategies for concept definition and validation on clinical physiological time series is needed to better understand how to extract clinically relevant information from the concept sensitivity scores.

Forecasting intraoperative hypotension during hepatobiliary surgery.

Prediction and avoidance of intraoperative hypotension (IOH) can lead to less postoperative morbidity. Machine learning (ML) is increasingly being applied to predict IOH. We hypothesize that incorporating demographic and physiological features in an ML model will improve the performance of IOH prediction. In addition, we added a "dial" feature to alter prediction performance. An ML prediction model was built based on a multivariate random forest (RF) trained algorithm using 13 physiologic time series and patient demographic data (age, sex, and BMI) for adult patients undergoing hepatobiliary surgery. A novel implementation was developed with an adjustable, multi-model voting (MMV) approach to improve performance in the challenging context of a dynamic, sliding window for which the propensity of data is normal (negative for IOH). The study cohort included 85% of subjects exhibiting at least one IOH event. Males constituted 70% of the cohort, median age was 55.8years, and median BMI was 27.7. The multivariate model yielded average AUC = 0.97 in the static context of a single prediction made up to 8min before a possible IOH event, and it outperformed a univariate model based on MAP-only (average AUC = 0.83). The MMV model demonstrated AUC = 0.96, PPV = 0.89, and NPV = 0.98 within the challenging context of a dynamic sliding window across 40min prior to a possible IOH event. We present a novel ML model to predict IOH with a distinctive "dial" on sensitivity and specificity to predict first IOH episode during liver resection surgeries.

Machine learning-based prediction of clinical outcomes after traumatic brain injury: Hidden information of early physiological time series.

To assess the predictive value of early-stage physiological time-series (PTS) data and non-interrogative electronic health record (EHR) signals, collected within 24 h of ICU admission, for traumatic brain injury (TBI) patient outcomes. Using data from TBI patients in the multi-center eICU database, we focused on in-hospital mortality, neurological status based on the Glasgow Coma Score (mGCS) motor subscore at discharge, and prolonged ICU stay (PLOS). Three machine learning (ML) models were developed, utilizing EHR features, PTS signals collected 24 h after ICU admission, and their combination. External validation was performed using the MIMIC III dataset, and interpretability was enhanced using the Shapley Additive Explanations (SHAP) algorithm. The analysis included 1085 TBI patients. Compared to individual models and existing scoring systems, the combination of EHR and PTS features demonstrated comparable or even superior performance in predicting in-hospital mortality (AUROC = 0.878), neurological outcomes (AUROC = 0.877), and PLOS (AUROC = 0.835). The model's performance was validated in the MIMIC III dataset, and SHAP algorithms identified six key intervention points for EHR features related to prognostic outcomes. Moreover, the EHR results (All AUROC >0.8) were translated into online tools for clinical use. Our study highlights the importance of early-stage PTS signals in predicting TBI patient outcomes. The integration of interpretable algorithms and simplified prediction tools can support treatment decision-making, contributing to the development of accurate prediction models and timely clinical intervention.

Exploring the Hidden Complexity: Entropy Analysis in Pulse Oximetry of Female Athletes.

This study examines the relationship between physiological complexity, as measured by Approximate Entropy (ApEn) and Sample Entropy (SampEn), and fitness levels in female athletes. Our focus is on their association with maximal oxygen consumption (VO2,max). Our findings reveal a complex relationship between entropy metrics and fitness levels, indicating that higher fitness typically, though not invariably, correlates with greater entropy in physiological time series data; however, this is not consistent for all individuals. For Heart Rate (HR), entropy measures suggest stable patterns across fitness categories, while pulse oximetry (SpO2) data shows greater variability. For instance, the medium fitness group displayed an ApEn(HR) = 0.57±0.13 with a coefficient of variation (CV) of 22.17 and ApEn(SpO2) = 0.96±0.49 with a CV of 46.08%, compared to the excellent fitness group with ApEn(HR) = 0.60±0.09 with a CV of 15.19% and ApEn(SpO2) =0.85±0.42 with a CV of 49.46%, suggesting broader physiological responses among more fit individuals. The larger standard deviations and CVs for SpO2 entropy may indicate the body's proficient oxygen utilization at higher levels of physical demand. Our findings advocate for combining entropy metrics with wearable sensor technology for improved biomedical analysis and personalized healthcare.

A novel neural network for improved in-hospital mortality prediction with irregular and incomplete multivariate data

Accurate estimation of in-hospital mortality based on patients’ physiological time series data improves the performance of the clinical decision support systems and assists hospital providers in allocating resources. In practice, the data quality issues of missing values are ubiquitous in electronic health records (EHRs). Since the vital signs are usually observed with irregular temporal intervals and different sampling rates, it is challenging to predict clinical outcomes with sparse and incomplete multivariate time series. We propose an auto-regressive recurrent neural network (RNN) based model, dubbed the bi-directional recursive encoder–decoder network (BiRED), to jointly perform data imputation and mortality prediction. To capture complex patterns of medical time sequences, a 2D cross-regression with an RNN unit (2DCR-RNN) and an imputation block with an RNN unit (IB-RNN) are designed as the recurrent component of the encoder and decoder, respectively. Furthermore, a state initialization method is proposed to alleviate errors accumulated in the generated sequence. The experimental results on two real EHR datasets show that our proposed method can predict hospital mortality with high AUC scores.

Physics-informed neural networks for modeling physiological time series for cuffless blood pressure estimation

The bold vision of AI-driven pervasive physiological monitoring, through the proliferation of off-the-shelf wearables that began a decade ago, has created immense opportunities to extract actionable information for precision medicine. These AI algorithms model input-output relationships of a system that, in many cases, exhibits complex nature and personalization requirements. A particular example is cuffless blood pressure estimation using wearable bioimpedance. However, these algorithms need training over significant amount of ground truth data. In the context of biomedical applications, collecting ground truth data, particularly at the personalized level is challenging, burdensome, and in some cases infeasible. Our objective is to establish physics-informed neural network (PINN) models for physiological time series data that would use minimal ground truth information to extract complex cardiovascular information. We achieve this by building Taylor’s approximation for gradually changing known cardiovascular relationships between input and output (e.g., sensor measurements to blood pressure) and incorporating this approximation into our proposed neural network training. The effectiveness of the framework is demonstrated through a case study: continuous cuffless BP estimation from time series bioimpedance data. We show that by using PINNs over the state-of-the-art time series models tested on the same datasets, we retain high correlations (systolic: 0.90, diastolic: 0.89) and low error (systolic: 1.3 ± 7.6 mmHg, diastolic: 0.6 ± 6.4 mmHg) while reducing the amount of ground truth training data on average by a factor of 15. This could be helpful in developing future AI algorithms to help interpret pervasive physiologic data using minimal amount of training data.

Identifying responders to elamipretide in Barth syndrome: Hierarchical clustering for time series data

BackgroundBarth syndrome (BTHS) is a rare genetic disease that is characterized by cardiomyopathy, skeletal myopathy, neutropenia, and growth abnormalities and often leads to death in childhood. Recently, elamipretide has been tested as a potential first disease-modifying drug. This study aimed to identify patients with BTHS who may respond to elamipretide, based on continuous physiological measurements acquired through wearable devices.ResultsData from a randomized, double-blind, placebo-controlled crossover trial of 12 patients with BTHS were used, including physiological time series data measured using a wearable device (heart rate, respiratory rate, activity, and posture) and functional scores. The latter included the 6-minute walk test (6MWT), Patient-Reported Outcomes Measurement Information System (PROMIS) fatigue score, SWAY Balance Mobile Application score (SWAY balance score), BTHS Symptom Assessment (BTHS-SA) Total Fatigue score, muscle strength by handheld dynamometry, 5 times sit-and-stand test (5XSST), and monolysocardiolipin to cardiolipin ratio (MLCL:CL). Groups were created through median split of the functional scores into “highest score” and “lowest score”, and “best response to elamipretide” and “worst response to elamipretide”. Agglomerative hierarchical clustering (AHC) models were implemented to assess whether physiological data could classify patients according to functional status and distinguish non-responders from responders to elamipretide. AHC models clustered patients according to their functional status with accuracies of 60–93%, with the greatest accuracies for 6MWT (93%), PROMIS (87%), and SWAY balance score (80%). Another set of AHC models clustered patients with respect to their response to treatment with elamipretide with perfect accuracy (all 100%).ConclusionsIn this proof-of-concept study, we demonstrated that continuously acquired physiological measurements from wearable devices can be used to predict functional status and response to treatment among patients with BTHS.

A pilot study to predict cardiac arrest in the pediatric intensive care unit

BackgroundCardiac arrest is a leading cause of mortality prior to discharge for children admitted to the pediatric intensive care unit. To address this problem, we used machine learning to predict cardiac arrest up to three hours in advance. MethodsOur data consists of 240 Hz ECG waveform data, 0.5 Hz physiological time series data, medications, and demographics from 1,145 patients in the pediatric intensive care unit at the Johns Hopkins Hospital, 15 of whom experienced a cardiac arrest. The data were divided into training, validating, and testing sets, and features were generated every five minutes. 23 heart rate variability (HRV) metrics were determined from ECG waveforms. 96 summary statistics were calculated for 12 vital signs, such as respiratory rate and blood pressure. Medications were classified into 42 therapeutic drug classes. Binary features were generated to indicate the administration of these different drugs. Next, six machine learning models were evaluated: logistic regression, support vector machine, random forest, XGBoost, LightGBM, and a soft voting ensemble. ResultsXGBoost performed the best, with 0.971 auROC, 0.797 auPRC, 99.5% sensitivity, and 69.6% specificity on an independent test set. ConclusionWe have created high-performing models that identify signatures of in-hospital cardiac arrest (IHCA) that may not be evident to clinicians. These signatures include a combination of heart rate variability metrics, vital signs data, and therapeutic drug classes. These machine learning models can predict IHCA up to three hours prior to onset with high performance, allowing clinicians to intervene earlier, improving patient outcomes.

Deep Learning-Based Health Management Model Application in Extreme Myopia Eye Vision Monitoring and Risk Prediction.

According to statistics released by the WHO, China has the highest prevalence of myopia in the world, with a frequency that is 1.5 times higher than the global average. Asians have the highest prevalence of myopia worldwide. The Ministry of Education and the State General Administration of Sports “2010 National Student Physical Fitness and Health Research Results” show that the incidence of poor vision among primary and secondary school students in China is 67.3%, and elementary school students' vision has decreased by 40.9%. Low vision among youth has become a major cause of affecting the quality of the population and improving national physical fitness; therefore, how to improve and enhance the vision level of youth has become a major issue for the government, sports, and educators face as a major issue. In order to address this issue, this research suggests a deep learning-based vision monitoring and risk prediction model for high myopia eyes and develops a deep artificial neural network that unsupervised learns essential characteristics of physiological time-series data.

Prediction of cybersickness in virtual environments using topological data analysis and machine learning

Recent significant progress in Virtual Reality (VR) applications and environments raised several challenges. They proved to have side effects on specific users, thus reducing the usability of the VR technology in some critical domains, such as flight and car simulators. One of the common side effects is cybersickness. Some significant commonly reported symptoms are nausea, oculomotor discomfort, and disorientation. To mitigate these symptoms and consequently improve the usability of VR systems, it is necessary to predict the incidence of cybersickness. This paper proposes a machine learning approach to VR’s cybersickness prediction based on physiological and subjective data. We investigated combinations of topological data analysis with a range of classifier algorithms and assessed classification performance. The highest performance of Topological Data Analysis (TDA) based methods was achieved in combination with SVMs with Gaussian RBF kernel, indicating that Gaussian RBF kernels provide embeddings of physiological time series data into spaces that are rich enough to capture the essential geometric features of this type of data. Comparing several combinations with feature descriptors for physiological time series, the performance of the TDA + SVM combination is in the top group, statistically being on par or outperforming more complex and less interpretable methods. Our results show that heart rate does not seem to correlate with cybersickness.

Novel Insights on Induced Sparsity in Multi-Time Attention Networks.

Current deep learning approaches for dealing with sparse irregularly sampled time-series data do not exploit the extent of sparsity of the input data. Our work is inspired by the sparse and irregularly sampled nature of physiological time series data in electronic health records. We explore the effect of inducing varying degrees of sparsity on the predictive performance of Multi-Time Attention Networks (mTAN) [1]. Our methodology is to induce sparsity by first sub-sampling the time-series before feeding it to the mTAN network. We conduct empirical experiments with sub-sampling ranging from 10 to 90 %. We investigate the performance of our methodology on the Human Activity dataset and Physionet 2012 mortality prediction task. Our results demonstrate that our proposed time-point sub-sampling coupled with mTAN improves the performance by 2 % on the Human Activity dataset with 80 % lesser time-points for training. On the Physionet dataset, our approach achieves comparable performance as baseline with 30 % lesser time-points. Our experiments reveal that time-series data could be further coarsely acquired when used in tandem with state-of-the-art networks capable of handling sparse data (mTAN). This could be of immense help for various applications where data acquisition and labeling is a significant challenge.

Computational signatures for post-cardiac arrest trajectory prediction: Importance of early physiological time series

BackgroundThere is an unmet need for timely and reliable prediction of post-cardiac arrest (CA) clinical trajectories. We hypothesized that physiological time series (PTS) data recorded on the first day of intensive care would contribute significantly to discrimination of outcomes at discharge. Patients and methodsAdult patients in the multicenter eICU database who were mechanically ventilated after resuscitation from out-of-hospital CA were included. Outcomes of interest were survival, neurological status based on Glasgow motor subscore (mGCS) and surrogate functional status based on discharge location (DL), at hospital discharge. Three machine learning predictive models were trained, one with features from the electronic health records (EHR), the second using features derived from PTS collected in the first 24 h after ICU admission (PTS24), and the third combining PTS24 and EHR. Model performances were compared, and the best performing model was externally validated in the MIMIC-III dataset. ResultsData from 2216 admissions were included in the analysis. Discrimination of prediction models combining EHR and PTS24 features was higher than models using either EHR or PTS24 for prediction of survival (AUROC 0.83, 0.82 and 0.79 respectively), neurological outcome (0.87, 0.86 and 0.79 respectively), and DL (0.80, 0.78 and 0.76 respectively). External validation in MIMIC-III (n = 86) produced similar model performance. Feature analysis suggested prognostic significance of previously unknown EHR and PTS24 variables. ConclusionThese results indicate that physiological data recorded in the early phase after CA resuscitation contain signatures that are linked to post-CA outcome. Additionally, they attest to the effectiveness of ML for post-CA predictive modeling.

Unsupervised Machine Learning Methods for Artifact Removal in Electrodermal Activity.

Artifact detection and removal is a crucial step in all data preprocessing pipelines for physiological time series data, especially when collected outside of controlled experimental settings. The fact that such artifact is often readily identifiable by eye suggests that unsupervised machine learning algorithms may be a promising option that do not require manually labeled training datasets. Existing methods are often heuristic-based, not generalizable, or developed for controlled experimental settings with less artifact. In this study, we test the ability of three such unsupervised learning algorithms, isolation forests, 1-class support vector machine, and K-nearest neighbor distance, to remove heavy cautery-related artifact from electrodermal activity (EDA) data collected while six subjects underwent surgery. We first defined 12 features for each halfsecond window as inputs to the unsupervised learning methods. For each subject, we compared the best performing unsupervised learning method to four other existing methods for EDA artifact removal. For all six subjects, the unsupervised learning method was the only one successful at fully removing the artifact. This approach can easily be expanded to other modalities of physiological data in complex settings.Clinical Relevance- Robust artifact detection methods allow for the use of diverse physiological data even in complex clinical settings to inform diagnostic and therapeutic decisions.

Digital signatures for early traumatic brain injury outcome prediction in the intensive care unit

Traumatic brain injury (TBI) is a leading neurological cause of death and disability across the world. Early characterization of TBI severity could provide a window for therapeutic intervention and contribute to improved outcome. We hypothesized that granular electronic health record data available in the first 24 h following admission to the intensive care unit (ICU) can be used to differentiate outcomes at discharge. Working from two ICU datasets we focused on patients with a primary admission diagnosis of TBI whose length of stay in ICU was ≥ 24 h (N = 1689 and 127). Features derived from clinical, laboratory, medication, and physiological time series data in the first 24 h after ICU admission were used to train elastic-net regularized Generalized Linear Models for the prediction of mortality and neurological function at ICU discharge. Model discrimination, determined by area under the receiver operating characteristic curve (AUC) analysis, was 0.903 and 0.874 for mortality and neurological function, respectively. Model performance was successfully validated in an external dataset (AUC 0.958 and 0.878 for mortality and neurological function, respectively). These results demonstrate that computational analysis of data routinely collected in the first 24 h after admission accurately and reliably predict discharge outcomes in ICU stratum TBI patients.

Machine Learning Analysis of Time-Dependent Features for Predicting Adverse Events During Hemodialysis Therapy: Model Development and Validation Study

BackgroundHemodialysis (HD) therapy is an indispensable tool used in critical care management. Patients undergoing HD are at risk for intradialytic adverse events, ranging from muscle cramps to cardiac arrest. So far, there is no effective HD device–integrated algorithm to assist medical staff in response to these adverse events a step earlier during HD.ObjectiveWe aimed to develop machine learning algorithms to predict intradialytic adverse events in an unbiased manner.MethodsThree-month dialysis and physiological time-series data were collected from all patients who underwent maintenance HD therapy at a tertiary care referral center. Dialysis data were collected automatically by HD devices, and physiological data were recorded by medical staff. Intradialytic adverse events were documented by medical staff according to patient complaints. Features extracted from the time series data sets by linear and differential analyses were used for machine learning to predict adverse events during HD.ResultsTime series dialysis data were collected during the 4-hour HD session in 108 patients who underwent maintenance HD therapy. There were a total of 4221 HD sessions, 406 of which involved at least one intradialytic adverse event. Models were built by classification algorithms and evaluated by four-fold cross-validation. The developed algorithm predicted overall intradialytic adverse events, with an area under the curve (AUC) of 0.83, sensitivity of 0.53, and specificity of 0.96. The algorithm also predicted muscle cramps, with an AUC of 0.85, and blood pressure elevation, with an AUC of 0.93. In addition, the model built based on ultrafiltration-unrelated features predicted all types of adverse events, with an AUC of 0.81, indicating that ultrafiltration-unrelated factors also contribute to the onset of adverse events.ConclusionsOur results demonstrated that algorithms combining linear and differential analyses with two-class classification machine learning can predict intradialytic adverse events in quasi-real time with high AUCs. Such a methodology implemented with local cloud computation and real-time optimization by personalized HD data could warn clinicians to take timely actions in advance.

Application of Deep Recurrent Q Network with Dueling Architecture for Optimal Sepsis Treatment Policy

Sepsis is one of the leading causes of mortality globally, and it costs billions of dollars annually. However, treating septic patients is currently highly challenging, and more research is needed into a general treatment method for sepsis. Therefore, in this work, we propose a reinforcement learning method for learning the optimal treatment strategies for septic patients. We model the patient physiological time series data as the input for a deep recurrent Q-network that learns reliable treatment policies. We evaluate our model using an off-policy evaluation method, and the experimental results indicate that it outperforms the physicians’ policy, reducing patient mortality up to 3.04%. Thus, our model can be used as a tool to reduce patient mortality by supporting clinicians in making dynamic decisions.

From sleep medicine to medicine during sleep–a clinical perspective

Objective. In this perspective paper, we aim to highlight the potential of sleep as an auspicious time for diagnosis, management and therapy of non-sleep-specific pathologies. Approach. Sleep has a profound influence on the physiology of body systems and biological processes. Molecular studies have shown circadian-regulated shifts in protein expression patterns across human tissues, further emphasizing the unique functional, behavioral and pharmacokinetic landscape of sleep. Thus, many pathological processes are also expected to exhibit sleep-specific manifestations. Modern advances in biosensor technologies have enabled remote, non-invasive recording of a growing number of physiologic parameters and biomarkers promoting the detection and study of such processes. Main results. Here, we introduce key clinical studies in selected medical fields, which leveraged novel technologies and the advantageous period of sleep to diagnose, monitor and treat pathologies. Studies demonstrate that sleep is an ideal time frame for the collection of long and clean physiological time series data which can then be analyzed using data-driven algorithms such as deep learning. Significance.This new paradigm proposes opportunities to further harness modern technologies to explore human health and disease during sleep and to advance the development of novel clinical applications – from sleep medicine to medicine during sleep.

PlethAugment: GAN-Based PPG Augmentation for Medical Diagnosis in Low-Resource Settings

The paucity of physiological time-series data collected from low-resource clinical settings limits the capabilities of modern machine learning algorithms in achieving high performance. Such performance is further hindered by class imbalance; datasets where a diagnosis is much more common than others. To overcome these two issues at low-cost while preserving privacy, data augmentation methods can be employed. In the time domain, the traditional method of time-warping could alter the underlying data distribution with detrimental consequences. This is prominent when dealing with physiological conditions that influence the frequency components of data. In this paper, we propose PlethAugment; three different conditional generative adversarial networks (CGANs) with an adapted diversity term for the generation of pathological photoplethysmogram (PPG) signals in order to boost medical classification performance. To evaluate and compare the GANs, we introduce a novel metric-agnostic method; the synthetic generalization curve. We validate this approach on two proprietary and two public datasets representing a diverse set of medical conditions. Compared to training on non-augmented class-balanced datasets, training on augmented datasets leads to an improvement of the AUROC by up to 29% when using cross validation. This illustrates the potential of the proposed CGANs to significantly improve classification performance.

Experimenting with Generative Adversarial Networks to Expand Sparse Physiological Time-Series Data.

Machine Learning research and its application have gained enormous relevance in recent years. Their usage in medical settings could support patients, increase patient safety and assist health professionals in various tasks. However, medical data is often sparse, which renders big data analytics methods like deep learning ineffective. Data synthesis helps to augment small data sets and potentially improves patient data integrity. The presented work illustrates how Generative Adversarial Networks can be applied specifically to small data sets for enlarging sparse data. Following a state-of-the-art analysis is conducted, experimental methods with such networks are documented, which have been applied to three different data sets. Results from all three sets are presented and take-away messages are summarized. Concluding, the results' quality and limitations of the work are discussed.

Dynamic and explainable machine learning prediction of mortality in patients in the intensive care unit: a retrospective study of high-frequency data in electronic patient records

Many mortality prediction models have been developed for patients in intensive care units (ICUs); most are based on data available at ICU admission. We investigated whether machine learning methods using analyses of time-series data improved mortality prognostication for patients in the ICU by providing real-time predictions of 90-day mortality. In addition, we examined to what extent such a dynamic model could be made interpretable by quantifying and visualising the features that drive the predictions at different timepoints. Based on the Simplified Acute Physiology Score (SAPS) III variables, we trained a machine learning model on longitudinal data from patients admitted to four ICUs in the Capital Region, Denmark, between 2011 and 2016. We included all patients older than 16 years of age, with an ICU stay lasting more than 1 h, and who had a Danish civil registration number to enable 90-day follow-up. We leveraged static data and physiological time-series data from electronic health records and the Danish National Patient Registry. A recurrent neural network was trained with a temporal resolution of 1 h. The model was internally validated using the holdout method with 20% of the training dataset and externally validated using previously unseen data from a fifth hospital in Denmark. Its performance was assessed with the Matthews correlation coefficient (MCC) and area under the receiver operating characteristic curve (AUROC) as metrics, using bootstrapping with 1000 samples with replacement to construct 95% CIs. A Shapley additive explanations algorithm was applied to the prediction model to obtain explanations of the features that drive patient-specific predictions, and the contributions of each of the 44 features in the model were analysed and compared with the variables in the original SAPS III model. From a dataset containing 15 615 ICU admissions of 12 616 patients, we included 14 190 admissions of 11 492 patients in our analysis. Overall, 90-day mortality was 33·1% (3802 patients). The deep learning model showed a predictive performance on the holdout testing dataset that improved over the timecourse of an ICU stay: MCC 0·29 (95% CI 0·25-0·33) and AUROC 0·73 (0·71-0·74) at admission, 0·43 (0·40-0·47) and 0·82 (0·80-0·84) after 24 h, 0·50 (0·46-0·53) and 0·85 (0·84-0·87) after 72 h, and 0·57 (0·54-0·60) and 0·88 (0·87-0·89) at the time of discharge. The model exhibited good calibration properties. These results were validated in an external validation cohort of 5827 patients with 6748 admissions: MCC 0·29 (95% CI 0·27-0·32) and AUROC 0·75 (0·73-0·76) at admission, 0·41 (0·39-0·44) and 0·80 (0·79-0·81) after 24 h, 0·46 (0·43-0·48) and 0·82 (0·81-0·83) after 72 h, and 0·47 (0·44-0·49) and 0·83 (0·82-0·84) at the time of discharge. The prediction of 90-day mortality improved with 1-h sampling intervals during the ICU stay. The dynamic risk prediction can also be explained for an individual patient, visualising the features contributing to the prediction at any point in time. This explanation allows the clinician to determine whether there are elements in the current patient state and care that are potentially actionable, thus making the model suitable for further validation as a clinical tool. Novo Nordisk Foundation and the Innovation Fund Denmark.